ByteDance’s billion-RMB bet on China’s first AI-native hospital

(By Caixin journalist Tang Hanyu)

While the global healthcare industry tentatively experiments with artificial intelligence (AI), one private hospital in Beijing is going all in. Its goal: to move beyond using AI as a mere physician’s assistant and instead install it as the brain of the entire institution.

Ground was broken nearly nine months ago in northeastern Beijing on a new campus for Arion Cancer Center, a massive project with an estimated investment of 6 billion RMB (US$872 million) to hold 800 beds. It will eventually expand beyond oncology into a comprehensive tertiary international hospital.

But its true ambition lies in its blueprint: becoming China’s first “AI-native” brick-and-mortar hospital.

Opened in 2021, Arion is the only private tertiary cancer hospital in Beijing, operating alongside state heavyweights like the Cancer Hospital of the Chinese Academy of Medical Sciences and Peking University Cancer Hospital.

Arion is part of the Beijing-based AmCare hospital group, which boasts the deep pockets of tech giant ByteDance Co. Ltd., owner of TikTok. In 2022, ByteDance, through its healthcare brand Xiaohe Health Technology (Bejing) Ltd., took full control of AmCare — originally known for its women’s and children’s hospitals. The high-premium acquisition, valued at 10 billion RMB, still holds the record for a private hospital transaction in China. The deal also gave ByteDance a 91% stake in Arion’s parent company, Hongda Arion.

Xu Zhonghuang, president of Arion, told Caixin that while public hospitals have an abundance of medical students, residents and fellows, human resources remain a major bottleneck for private hospitals. However, Xu said, “A good large language model is already easily as capable as a senior resident.”

Xu hopes AI will first prove “capable” by taking over repetitive, low-efficiency tasks. In the future, he expects it to become “smart” — offering insights and predictions for treatment techniques, methodologies and drugs where human capacity falls short. “The ultimate goal of building an AI hospital is to explore the future of healthcare,” he said.

Xu envisions AI agents embedded throughout the hospital’s architecture, medical equipment and information networks, covering all crucial operations to form a “collaborative ecosystem” with human staff.

If all goes smoothly, Arion expects the new hospital to open by the end of 2029.

The medical AI landscape in China is characterised by widespread experimentation, but holistic integration is rare, explained Hu Sanduo, IT and data director at Arion. Usually, AI tools fail to communicate with existing hospital tech systems, stunting collaboration. Furthermore, the lack of platforms to optimise models and govern data prevents the system from feeding improvements back into the AI.

“These highly fragmented applications make it difficult to establish comprehensive context for the AI, and there is no interaction mechanism between apps or between apps and human employees,” Hu told Caixin.

As a result, Arion opted for a holistic design. This addresses three fundamental contradictions: data must flow, systems must collaborate and models must grow.

The vision relies on a specialised medical cloud powered by Volcengine, ByteDance’s cloud platform, to serve as the hospital’s bedrock. On top of that, all hospital operations and data capabilities will be deployed as AI agents.

This happens in two steps: first, creating a unified “data foundation” to connect underlying operational systems; second, using Volcengine’s large language models to manage and orchestrate the agents, essentially minting “digital employees”.

These digital workers fall into three categories. Data agents operate in the background, acting as simultaneous interpreters to ensure different systems understand one another. Role agents act as physician assistants or medical residents. Finally, business agents handle specific, singular tasks. The data these agents collect is then fed back into the foundation, enriching the system’s contextual awareness.

“Reconstructing the entire hospital information system through these agent-based digital employees is our core objective,” Hu said.

This AI-native operational system will ideally span five key domains — routine health management, clinical operations, clinical research, medical education and hospital administration — covering over 100 core scenarios.

Armed with rich context, agents can move beyond generic advice, offering highly targeted patient recommendations or holistic financial analyses for hospital management.

A tool currently in development — one that doctors are eagerly anticipating — is called PDSP.

In Arion’s system, it appears as a dashboard of parallel modules. It consolidates a patient’s timeline of tests, lab results, pathology, diagnoses, medical records and nursing notes. But it also searches PubMed documentation and cross-references Arion’s historical multidisciplinary consultation database. It then feeds this localised expertise to a general large language model previously trained on clinical guidelines and expert consensus.

“When we develop a treatment plan, we are referencing standards, the latest medical literature, the patient’s personal data, and, crucially, past discussions on similar cases,” said Pan Yong, a deputy chief physician at Arion’s breast centre. “This makes the process much more efficient and personalised.” Hu envisions a scenario where a doctor drops a prompt into a group chat and digital “colleagues” jump into action: the pharmacy prepares the medication, a nurse preps the injection and someone else sends the patient care instructions.

That seamless collaboration does not exist yet. Currently, however, AI at Arion acts primarily as a data aggregator. Any generated outputs undergo dual verification by both AI and a human. “We are still far off. People are still operating the systems,” Hu said. “After all, this is a five-year plan, and we have to keep strengthening it.”

While AI still functions as an assistant in these one-on-one scenarios, Hu noted that the long-term goal is to build a team of humans collaborating with a team of AI agents.

To fully establish the collaborative network, Arion should first maximise the store of data.

Medical data fed to AI generally falls into two camps. The first is out-of-hospital health management data, including social determinants, follow-up consultations and wearable device metrics. The second is in-hospital clinical data, ranging from unstructured clinical notes to structured data, medical imaging and genomics.

In traditional hospitals, this in-house data is scattered across departments and software systems, and high-quality data is scarce. Hu noted that tracking crucial patient data that historically lived in manila folders is painstakingly difficult.

Knowing this, the Arion team made data structuring a priority when building their own IT systems during the hospital’s early planning stages in 2019.

Arion’s strategy was to purchase an open-architecture, mid-sized hospital information system and overhaul it piece by piece. Their resulting proprietary system organises patient data chronologically from the start. This “releases the patient from the manila folder” and makes it easier for algorithms to deduce logical relationships between events based on time.

As a for-profit facility, Arion has yet to reach a break-even point. “Our speed to break even is actually much faster than average, but the investment is indeed massive,” Xu said, adding that ByteDance continues to underwrite the vision, so long as the demands are sensible.

“We aren’t taking shortcuts to expand scale for profit. We are elevating our standards and our technology. This aligns with ByteDance’s direction and keeps us from being heavily burdened,” the president said.

This article was first published by Caixin Global as “In Depth: ByteDance’s Billion-Yuan Bet on China’s First AI-Native Hospital”. Caixin Global is one of the most respected sources for macroeconomic, financial and business news and information about China.

Related: China’s tech giants burn cash to try to dominate AI healthcare | Pudu Robotics founder Zhang Tao: Humanoid robots will be bigger than cars